[Scummvm-devel] Predictive input code

Kostas Nakos knakos at phys.uoa.gr
Mon Apr 7 16:32:31 CEST 2008


Hello,

On 04/04/2008 03:45 PM, Max Horn wrote:
> At one point, I had rewritten that code to use a trie for lookup. This was
> scractched and replaced, though, because it caused super-slow loadup times
> for some backends which had an inferior malloc implementation.

Using a trie is certainly the correct way to go about this. As it 
stands, the prediction jumps widely between words as you type. It's not 
that noticeable, as it converges "sanely" as one types - but 
nevertheless it can be done better with a trie.

Note also that on WinCE, using the former code, slightly increased load 
times were observed but mainly mem exhaustion (fragmentation, ..., who 
knows) was prominent.

> So, I would like to try to revert the code back to a trie based algorithm
> (based on my old code in there, with some refinements, which should cut
> down the mallocs). Before I do this, I want to know whether there are
> other plans for the predicitve code (is *anybody* doing anything with
> this?). And whether the porters whose ports suffered from the code would
> be willing to test patches by me for speed regressions before I put them
> into SVN, in a timely fashion (I don't want my patch to bitrot for 6
> months, but I also don't want to upset porters by just committing it).

If you want to take a hack at it feel free, I am not intending to touch 
that code. Also I can promptly test any patches for you.

Best Regards,
Kostas




More information about the Scummvm-devel mailing list